基于GMM的说话人识别算法的研究与应用  

Research and Application of Speaker Recognition Algorithm Based on GMM

作　　者：吴慧玲[1] 杜成东 毛鹤[1] 

机构地区：[1]苏州大学电子信息学院,苏州215006

出　　处：《现代计算机（中旬刊）》2014年第5期31-35,共5页Modern Computer

基　　金：苏州大学大学生创新训练计划项目(No.2012yb011);省级大学生创新训练项目(No.2012yb011)

摘　　要：说话人识别是根据检测到的语音进行说话人身份的认证,是将待识别语音与数据库中的说话人语音进行匹配的过程。设计基于高斯混合模型(GMM)说话人识别系统,提取输入语音的Mel倒谱系数作为观察向量,用GMM算法进行说话人语音模型训练和识别。同时设计基于TMS320DM3730 DSP的嵌入式硬件平台,并在该平台上实现所设计的说话人识别系统。为进行性能测试,自行录制相应的语音材料库,录音的人数为38人,其中男19人,女19人。经测试表明,在正常环境下,该设计的说话人系统识别率可达到95%以上。Speaker identification recognizes the speaker＇s ID according to the acquired speech which is realized by matching the extracted charac- teristics of the input speech with their counterparts saved in the speech lib. Adopts a GMM to implement the speaker identification, where selects Mel-frequency eepstrum coefficients of the speech frame are selected as the observation vectors. Also designs an embedded hard- ware platform based on TMS320DM3730 DSP to realize the GMM based identification algorithm. Creates a voice database including 19 men＇s voice segments and 19 women＇s voice segments for the training and performance testing. The testing results show that under normal environment, the identification accuracy can be above 95%.

关 键 词：说话人识别 GMM MEL倒谱系数 DM3730 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]